Claude Code compares conflicting MRI interpretations
Antoine used Claude Code to analyze a shoulder DICOM export and compare reports. Two generated PDFs disputed the clinician’s interpretation, leaving the diagnosis unresolved.
Original prompt
Shared by the source author. See source notes for context.
right shoulder pain for 2–3 weeks
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Antoine
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Two linked PDF reports were generated and disagreed with the clinician about a tendon tear; the source leaves diagnostic correctness unresolved.
- Human involvement
- Antoine supplied records and context, compared contradictory outputs and retained uncertainty. No later clinician-confirmed resolution is documented.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Provide the DICOM archive and allow package installation and code execution.
- Ask Claude Code to plan the analysis and generate an initial PDF report.
- Compare its interpretation with the clinical report and supply additional context.
Antoine used Claude Code to inspect a shoulder MRI after a clinical appointment left him uncertain about the proposed treatment. He supplied a DICOM export of several hundred files, about 266 MB, and allowed Claude Code to install analysis packages and run code. The model used was Opus 4.8.
The first run received only a short symptom description. After about an hour, it produced a PDF report that described an intact tendon, conflicting with the clinician's report of a substantial partial tear. Antoine then supplied the clinical report and additional context and requested a comparison. The second run used multiple subagents and generated an arbitration PDF, again favoring the model's interpretation on the disputed tear.
The useful documented outcome is the production of two inspectable reports and a clear disagreement, not a medically verified second diagnosis. Antoine links both PDFs and says that some differences between reports could not be resolved. The article ends with his uncertainty about whom to trust; it does not provide a later clinical adjudication.
This case shows how a coding agent can process an imaging archive and assemble a long analysis while leaving a consequential interpretation unverified. The initial agent also received less context than the clinician. Package execution, time spent and agreement among subagents do not establish diagnostic accuracy. The source is a personal experiment, and its conclusions require clinical review.
Prompt
Original prompt
Source: Blog · Antoine
Frequently asked questions
Was the model’s diagnosis clinically confirmed?
No. The author reports conflicting interpretations and uncertainty, without a later clinical adjudication.
What did Claude Code receive initially?
A DICOM export of several hundred files and a short shoulder-pain description, which the author says was less context than the doctors had.
What changed in the second run?
The author supplied the clinical report and more context, then requested a comparison that used multiple subagents and produced another PDF.
More details
Remaining steps
- Request a second analysis using subagents and inspect unresolved disagreements.
About this account
Claude Code processed imaging files, ran analysis code and created reports; the comparison run used subagents without individual human prompts for each analysis.
Firsthand source opened over HTTP and read on 2026-10-15. Publication date comes from the source page. Outcomes are author-reported and were not independently reproduced. Personal medical experiment; clinical correctness unresolved; initial context differed from clinician context; no treatment advice or model interpretation endorsed. Original prompt box contains a short excerpt of the author's published instructions.
Credit: Antoine
Source published
Added to Agents In Use
Source checked
Tools and topics
Claude Code · DICOM files · Analysis packages · Subagents · PDF reports
Imaging analysis · Conflicting reports · Unresolved outcome
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.